Abstract
The Bayesian brain theory is currently having a large influence in experimental psychology and cognitive (neuro)science. Thought to originate from the 1867 von Helmholtz’s notion of “unconscious inference”, this theory posits that there are three basic elements to perform perceptual inference: 1) sampling of sensory evidence, 2) extraction of statistical regularities to create models about the environment, and 3) exploitation of such predictive models. These three building blocks do not work as a serial chain but interact dynamically to optimize the decision-making process (e.g., if sensory evidence points in a direction opposite to what predicted, models must be readjusted through new learning). The current literature shows a plethora of behavioral evidence about the potential role of predictive coding in perception, while there are very little systematic attempts to understand their precise psychophysiological mechanisms, in particular within an oscillatory framework. This is of particular relevance if we consider that the constituent blocks of Bayesian inference and their dynamic integration appear to be altered in the autistic spectrum disorder (ASD) and schizotypal spectrum disorder (SSD) continuum: while sensory evidence would be predominant in ASD-like perception, SSD would rely excessively on priors to guide perception and behavior. Notably, disruption in oscillatory activity is evident in both ASD and SSD, but their influence in the context of predictive perception remains unknown. In the present project we propose to conduct the first systematic investigation of the psychophysiological mechanisms underlying human predictive processes, with a specific focus on rhythmic oscillatory activity, and test the interindividual differences as a function of ASD and SSD traits. Specifically, the project delineated the constituent building blocks of predictive perception and their dynamics interplay to reach a deep understanding of: (I) how sensory evidence is weighted in perceptual decisions and how it can give rise to low-level priors, (II) the temporal dynamics underlying statistical priors learning and the interplay between low-level priors and implicitly learned prior models, (III) the impact that explicit high-level priors have on perception and its interplay with low-level priors. The project benefited from a large data collection (N=420) and from complementary and synergistic expertise of two research groups (UNIBO and UNITN) in methods including electroencephalography, psychophysics, computational modeling, as well as complementary expertise in ASD and SSD. The experimental settings we propose were tuned to create innovative contexts able to isolate the building blocks of Bayesian inference (sensory evidence sampling, prior formation and exploitation), but where also dynamic decision-making and flexible response are required, mimicking more ecological contexts and allowing to shed lights on fundamental individual differences in the predictive brain.
Results achieved
The project delivered the first systematic investigation of the psychophysiological – and specifically oscillatory – mechanisms underlying predictive coding in human perception, within the "Bayesian brain" framework. By combining psychophysics, electroencephalography and computational modelling in a large general-population sample, it decomposed perceptual inference into three partially dissociable components – sensory evidence weighting, implicit prior learning, and explicit prior exploitation – showing that each is supported by distinct oscillatory signatures and that these components are differently modulated by individual characteristics along the autism-schizophrenia continuum. First, using a perceptual decision-making task based on dynamic variations in motion coherence, the project tested how changes in evidence introduced at different processing stages shape choice. An increase in coherence (facilitating pulse) improved performance, whereas incongruent evidence worsened it; crucially, the impact was greater when the change occurred at earlier stages of accumulation, confirming that the system does not weight information uniformly over time but privileges early evidence. At the neural level, the effect was associated with alpha/low-beta desynchronisation (about 7–17 Hz) over centro-parietal electrodes, more pronounced for the facilitating than for the interfering pulse in a window around 500–850 ms. Trait analyses indicated a closer link with the autistic-spectrum dimension (AQ) than with the schizotypal one (SPQ), suggesting a more selective role of autistic traits in how sensory evidence is incorporated into the decision. Secondly, we investigated the neural underpinning of prior learning. In a feedback-based detection task, participants had to infer, trial by trial, the most advantageous decision strategy associated with two cue-defined contexts, without explicit instructions. Perceptual sensitivity (d') increased with practice, but the critical signal was the decision criterion, which participants progressively differentiated according to context – more conservative when stimulus-absent responses were rewarded, more liberal when stimulus-present responses were rewarded; the total score increased over time and was correlated with the magnitude of this modulation. At the electrophysiological level a sequential theta–alpha architecture emerged: feedback-related midfrontal theta was strongest in the early phase, consistent with the updating of the internal model during rule acquisition, whereas pre-stimulus posterior alpha became relevant later, when the learned strategy was implemented in anticipation of the stimulus. Distinguishing good from poor learners, the former showed greater criterion adaptation, better point accumulation, stronger early theta and, in later phases, a context-sensitive alpha modulation (greater alpha-beta desynchronisation in the liberal condition); both indices were associated with the degree of behavioural adjustment. The two groups, however, did not differ in AQ/SPQ scores: in this non-clinical sample the efficiency of implicit learning depended on performance-based rather than dispositional factors. Finally, we investigated the neuro-behavioural underpinning of explicit prior usage. Using a probabilistic detection task – in which an explicit cue signalled the probability of target occurrence (high, low or neutral) – explicit information did not alter perceptual sensitivity but modulated the criterion, made more liberal by high-probability cues and more conservative by low-probability ones. This modulation was tracked by the amplitude of pre-stimulus posterior alpha, with a spatially localised effect over posterior regions contralateral to the stimulus; distinguishing "believers" from "empiricists" (strong vs weak alpha modulators), the former strongly bound the criterion to the prior, the latter much less. Crucially, predictive style varied dimensionally along the continuum: proximity to the schizotypal pole was associated with greater reliance on priors, proximity to the autistic pole with a greater weight of sensory evidence, and posterior alpha amplitude partially mediated the relationship between position on the continuum and the extent of criterion adjustment. Overall, the project provides a unified account of the oscillatory mechanisms of predictive perception – alpha/low-beta for evidence weighting and explicit prior exploitation, theta for learning and monitoring – and a dimensional framework, beyond categorical diagnostic boundaries, for understanding different predictive-processing styles and informing future translational work on neurodevelopmental and schizophrenia-spectrum conditions. Results were presented at several national and international conferences between 2022 and 2025. The project, coordinated by the University of Bologna (P.I. Vincenzo Romei) in partnership with the University of Trento (Co-P.I. Luca Ronconi), was funded by the European Union – Next Generation EU under the NRRP (PNRR). The results and the theoretical framework of the project have fed into several international publications: • Tarasi, L., Alamia, A., & Romei, V. (2026). Backward alpha band oscillations shape perceptual bias under probabilistic cues. Communications Biology. • Tarasi, L., Alamia, A., & Romei, V. (2025). Perceptual bias in motion discrimination is related to asymmetric interhemispheric alpha traveling waves. Advanced Science, 12(41), e14623. • Frisoni, M., Tarasi, L., Borgomaneri, S., & Romei, V. (2025). The relationship between individual alpha frequency and time perception: Testing the internal clock versus the sampling rate hypothesis. Cortex. • Trajkovic, J., Ricci, G., Pirazzini, G., Tarasi, L., Di Gregorio, F., Magosso, E., Ursino, M. & Romei, V. (2025). Aberrant Functional Connectivity and Brain Network Organization in High-Schizotypy Individuals: An Electroencephalography Study. Schizophrenia Bulletin, 51(5), 1266-1281. • Tarasi L, Covelli M, Tabarelli de Fatis C, Romei V. (2025) Prior Information Shapes Perceptual Confidence. Journal of Cognition 2025 Jan 7;8(1):11. doi: 10.5334/joc.417. PMID: 39822242; PMCID: PMC11736391.Dettagli del progetto
Responsabile scientifico: Vincenzo Romei
Strutture Unibo coinvolte:
Dipartimento di Psicologia "Renzo Canestrari"
Coordinatore:
ALMA MATER STUDIORUM - Università di Bologna(Italy)
Contributo totale di progetto: Euro (EUR) 204.772,00
Contributo totale Unibo: Euro (EUR) 103.090,00
Durata del progetto in mesi: 24
Data di inizio
05/10/2023
Data di fine:
28/02/2026
